How to add AI to an app you already have
The five steps from an app with no AI to one where users get work done by asking, without rebuilding anything.
Quick answer
To add AI to an existing app, start from a job your users already struggle with, not from the model. Pick the shape that fits that job: an AI feature inside one screen, an assistant that answers questions, or an agent that does multi-step work for the user through the API your app already has. Connect it as the signed-in user, try it on test data, ship it to a few users, and let what they ask for decide what comes next.
Most apps that add AI start in the wrong place. Someone picks a model, wires up a chat box, and then looks for something for it to do. Users open it once, ask a question the product could already answer, and never open it again.
The apps where AI sticks work the other way round. They start from a job users already find hard, and add exactly enough AI to make that job easy. Here is that process, in five steps, for an app you already have.
1. Find the job users struggle with
You are looking for work that takes your users longer than it should. The signs are usually already in front of you:
- Requests for features you already have. The feature exists; the path to it does not. (More on this in users keep asking for features you already have.)
- A workflow people start and do not finish. A setup with six steps, a report that needs three filters, an import that fails on the second screen.
- The same "how do I" on every sales call. If prospects ask it, users who never got a call are stuck on it too.
Write the job down the way a user would say it: "move all my Friday meetings to next week and let the guests know", not "bulk reschedule". That sentence is what you will test against later.
2. Pick the shape that fits the job
There are three ways AI shows up inside an app, and the job decides which one you need.
An AI feature inside one screen. Summarise this thread, draft this reply, sort these records. One button, one model call, one result. Cheap to build and easy to explain, and the right choice when the job lives on a single screen.
An assistant that answers. Users ask, it replies from what you have written down. Right when the job is finding information that already exists.
An agent that does the work. Users say what they want, and it carries out the steps for them across the product: finds the meetings, moves them, sends the messages. Right when the job spans several screens, which is where users get stuck most.
If you are unsure which you need, the questions in the best ways to add an AI assistant to a SaaS settle it.
3. Connect it to what you already have
You do not need to rebuild anything. Your app already has an API: the one your own screens call. An AI feature calls a model from your backend and returns the result to that screen. An agent should use the same API your screens do, as the user who is signed in, so it can do exactly what that user could do by hand and nothing more.
That last part matters more than it looks. An agent running on an admin key can do things the user never could, and when something goes wrong your records say "the AI did it" rather than who asked. An agent acting as the user inherits every permission you already got right.
How much there is to connect varies widely. The Verb installs I have set up range from 29 actions in a law firm's practice management product to 68 in a marketplace admin panel; the live scheduling demo has 35.
4. Try it before your users do
Run it against test data first, with the sentences you wrote down in step 1. Then try the requests that should fail: something the user is not allowed to do, a date that does not exist, a request that is half in another language. Anything that changes data should show the user what it is about to do and wait for a yes. A separate test copy of your app, on your own computer, is the cheapest place to find out what breaks.
5. Ship it small, then read what people ask
Turn it on for a slice of users, and read every request for the first couple of weeks. The requests it completed tell you which jobs were hidden. The ones it could not complete tell you what users expected your product to do, in their own words, which is a better roadmap than any vote count.
How long it takes
An AI feature in one screen is days. An agent you build yourself is weeks to a first demo and months to something you would put in front of every user, because most of the work comes after the model call: requests that go wrong, permissions, testing, and keeping it working as the product changes. Why most in-app AI agents never ship covers where that time goes.
If you would rather have the agent than build it, that is what Verb is for. You describe what it may do in your product, try it on a test copy, and add one script tag. It is integrated in under an hour in most cases, and the docs walk through each step.
Frequently asked questions
How do I integrate AI into my app without rebuilding it?
Use what the app already has. A single AI feature is one call to a model from your backend. An agent should work through the same API your own screens use, as the signed-in user, so nothing is rebuilt and nothing new needs permissions. With Verb you describe those actions once and add one script tag, and it is integrated in under an hour in most cases.
Do I need an AI team to add AI to my app?
Not for an AI feature inside one screen, and not for an agent you embed rather than build. You need one if the agent itself is your product and you want to own every part of it, because most of the work comes after the first demo.
How much does it cost to add AI to an app?
The model is rarely the big cost. What costs most is engineering time after the first demo: handling the requests that go wrong, testing against real data, and keeping it working as the product changes. Measure the time it saves your users before deciding what it is worth.
Should I add a chatbot or an AI agent to my app?
Look at what users ask for. If the answer is already written down, an assistant that answers is enough. If they want something done that takes several screens, such as setting up a recurring schedule or exporting a filtered list, an agent that does it for them is the better fit.
Vishnu, founder of AskVerb @itsvishnups
Building the AI agent that lives inside a SaaS product and does the work its users ask for. Writing down what I learn along the way, including the parts I got wrong first.